ICYMI: Meta's Muse AI Agent Uses Your Chats with It to Train Meta's AI by Default

Meta’s Muse AI agent uses conversations to help train Meta’s AI models by default, according to a privacy warning published by MEGA. Users who do not want their chats included must find and disable a model-improvement setting themselves.

The warning also says Muse nudges users to share bank account, email, and passport information. Meta says it strips key personal identifiers before sending data into training, but that safeguard does not resolve every privacy or cybersecurity concern raised by collecting sensitive information in the first place.

Privacy-conscious users can open Settings, select Data controls, and turn off “Help improve our AI models.” The setting leaves other questions unanswered, including whether earlier chats remain eligible for training and how long Meta retains the original information.

Muse Makes AI Training the Default

MEGA’s post says Muse users are automatically enrolled in model improvement unless they opt out. Citing Wired, it characterizes users as “unpaid AI trainers,” a pointed description of the value companies can extract from conversations at scale.

The post also says Muse reached No. 1 on Apple’s App Store. The supplied evidence does not provide important context for that ranking, such as the country, chart category, or date. Even so, rapid adoption would increase the number of people affected by a default that many could overlook.

Model training does not necessarily mean Meta employees manually read every conversation. It generally refers to collected data being used in systems that evaluate, fine-tune, or otherwise improve AI models. The available evidence does not establish the exact Muse process, which models receive the data, or how conversations are selected.

Users can therefore open Muse expecting an AI assistant and, unless they inspect its settings, also contribute their conversations to future model development.

How to Opt Out of Muse AI Training

The reported opt-out process takes three steps:

1. Open Settings in Muse.
2. Select Data controls.
3. Turn off “Help improve our AI models.”

Afterward, users should return to the Data controls screen and confirm that the switch remains disabled. If the option is missing, they should check Muse’s current in-app privacy information instead of assuming training is already off.

The wording suggests this control governs the secondary use of conversations for model improvement. It does not promise that Muse will stop processing a chat entirely. The agent still needs to process user input to generate a response, and the switch does not explain operational retention, security logs, account history, or legal obligations.

The supplied information also does not establish that opting out deletes previously collected data. Anyone seeking to remove earlier conversations should look separately for chat-history and account-deletion controls. Meta needs to clarify whether disabling training applies retroactively or only to data submitted after the setting changes.

Removing Identifiers Does Not Make a Chat Private

Meta says it strips key personal identifiers before using Muse data for training. That can reduce risk, but “key personal identifiers” is not a precise technical inventory.

Direct identifiers include fields such as a name, email address, passport number, or bank account number. A combination of other details can also identify someone: their employer, location, job title, travel dates, medical history, or an unusual personal event. Removing a name does not automatically make the remaining text anonymous.

Timing also matters. Stripping identifiers before training means Meta’s systems receive the original information before applying the removal process. By itself, the claim does not answer:

• How long the original chat is stored
• Whether uploaded images and attachments are covered
• Which identifiers the filters recognize
• Whether structured numbers are removed from free-form text
• Whether a person or automated system can review flagged conversations
• Whether Meta has independently tested the process for reidentification risks

There is no evidence here that Meta’s filters have failed, nor is there independent evidence showing how comprehensively they handle passports, financial details, attachments, or indirect identifiers. De-identification lowers some risks, but it does not guarantee anonymity or reduce the amount of data collected.

Sensitive Prompts Create a Cybersecurity Risk Too

The report says Muse nudges users to provide bank account, email, and passport information. A nudge is not necessarily a requirement, and the available evidence does not show which tasks produce these prompts. Users should still pause before entering information that could enable fraud, identity theft, or account recovery.

AI conversations can feel informal even when the requested data is highly sensitive. Unless the product clearly explains why information is required, how it will be protected, and when it will be deleted, users should not treat a chat field like a dedicated identity-verification or banking portal.

A few precautions can reduce unnecessary exposure:

• Never provide passwords, PINs, one-time security codes, recovery phrases, or complete payment-card credentials.
• Share only the minimum fields required for a specific task.
• Redact unrelated information from identity documents before uploading them, when redaction is permitted.
• Review any account permissions Muse requests and deny access that the task does not need.
• Revoke unused connections after completing a task.
• Change a credential immediately if it was accidentally pasted into a conversation.

AI agents deserve additional scrutiny when they can interact with email, financial services, files, or other accounts. Because the supplied evidence does not establish Muse’s exact integrations, users should review the permissions presented by the app instead of assuming broad access or complete isolation.

The Default Setting Is the Real Point of Dispute

Users face two separate decisions: whether Muse should process information for a requested task and whether that information should help train Meta’s future AI models.

An opt-out design combines those decisions until someone finds the relevant setting. The label “Help improve our AI models” describes a potential benefit to Meta and its products, but it does not tell users which conversations qualify, how long the training copy persists, or which models receive it.

An opt-in approach would present the choice before a conversation became eligible for training. People could then use the agent without first discovering that their data serves another purpose.

Muse’s reported App Store position makes the design more consequential, though popularity does not settle the privacy question. A widely used product should have clearer consent controls, particularly if it encourages people to supply identity and financial information.

Meta Still Needs to Explain the Data Flow